The artificial intelligence ecosystem is evolving at an exhilarating pace, with demands for more powerful and efficient AI models driving the wheels of AI innovation. At the forefront of this revolution stands NVIDIA, constantly pushing the limits of AI possibilities.
Recent data released on June 4 by MLCommons, a non-profit organization dedicated to AI performance evaluations, showed that NVIDIA’S chips, especially those built with Blackwell architecture, are delivering significant gains in AI performance by outperforming their predecessors and setting a new bar in the AI industry.
The Gold Standard: Understanding MLPerf Training Benchmarks
MLCommons developed a benchmark, MLPerf, designed by a consortium of over 125 AI leaders from academics, research labs, and industry to provide unbiased evaluations of training and inference performance for hardware, software, and services for both training and inference tasks, which is always adapting to the latest advancements in AI.
MLPerf is widely recognized as the industry’s most trusted and rigorous benchmark for evaluating AI performance. The range of tests it provides measures how quickly a platform can train various AI models to predetermine its quality thresholds, encompassing diverse workloads from image recognition and object detection to natural language processing and, majorly, large language model pre-training and fine-tuning.
The MLPert Training v5.5 suite introduced a new and more demanding benchmark: Llama 3.1 405B pretraining. This new Llama model, which serves as a representative for the current state-of-the-art LLMs, has a whopping 405 billion parameters, which serves as a true stress test for modern AI hardware and software stacks needed to keep up with the escalating demands of training the next-generation AI.
MLPerf Training v5.0: Blackwell Dominates Across the Board
The recent MLCommons training benchmarks provide compelling, empirical evidence of Blackwell’s supremacy. The results show the Blackwell chips, on a per-chip basis, delivered 2.6X faster than the previous generation Hopper chips, which had 80 billion transistors used for training large AI systems. In a remarkable demonstration of its prowess, a cluster of 2,496 Blackwell GPUs, part of the NVIDIA DGX GB200 NVL72 system, completed the Llama 3.1 405B pre-training benchmark in 27 minutes.
Its predecessor, Hopper, would have required over three times as many units to scale this task for that scale of time. This shows how much more efficient and enhanced Blackwell architectural advancements are, with its high-density liquid-cooled rack and fifth-generation NVIDIA NVLink and NVLink Switch interconnect technologies for scale-up.
Apart from processing speed, the MLCommons results also highlighted NVIDIA’s exceptional scaling efficiency. In the process of expanding from 512 to 2,496 GPUs, NVIDIA’s GB200 NVL72 system demonstrated a 90% strong scaling efficiency.
Dominance Across Diverse AI Workloads
The NVIDIA AI platform delivered the highest performance at scale on all seven benchmarks in the MLPerf Training v5.0 suite. They included:
- LLM Pre-Training (Llama 3.1 405B)—trained within 20.8 minutes
- LLM Fine-Tuning (Llama 2 70B-LoRA) – trained within 0.56 minutes
- Text-to-Image (Stable Diffusion v2)—trained within 1.4 minutes
- Graph Neural Network (R-GAT)—trained within 0.84 minutes
- Recommender (DLRM-DCNv2)—trained within 0.7 minutes
- Natural Language Processing (BERT)—trained within 0.3 minutes
- Object Detection (RetinaNet)—trained within 1.4 minutes.
Implications for the Future of AI
- Faster AI progress: with AI training being achieved at faster speeds, researchers and AI developers can try out more ideas and improve their models quicker than before, signifying rapid AI advancement.
- More People X Can Work with Big AI: As the training process becomes more efficient, it will be easier for more organizations to work with large-scale AI models like Llama 3.1.
- NVIDIA Stays Ahead: The MLPerf Training v5.0 solidifies NVIDIA’s position as the undisputed leader in AI training hardware, which is vital for the company as demand for its AI technology continues to skyrocket.
- Emphasis on Interconnects and Software: The amazing scaling ability NVIDIA achieved shows how important superfast connections between chips (like NVLink) and smart software (like NVIDIA’s CUDA-X libraries) are for maximizing the hardware.
As AI models become more complex and large, being able to train them in a quick and efficient manner is paramount. NVIDIA’s breakthrough at the MLPerf v5.0 is not just about faster chips but the potential future for AI chips and how they will change the future of the AI landscape for years to come.
